On November 30, 2017 7:14:12 PM GMT+01:00, Jan Hubicka <hubi...@ucw.cz> wrote: >> >> I have benchmarked this quite thoroughly while switching the defaults >for Core >> and Zen. Disabling alignment is pretty much consistent for >specfp/int for >> 2000,2006 and 2017 (last is Zen only) and quite consistent one. I >will analyze >> fma3d. It did not show with the alignment change alone, so it may be >related >> to other costs or just bad luck. > >I have opened https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83232 for >that. >The problem is that SLP gives up vectorization when it sees vectorized >used >in the same basic block. Preivously there was alignment prologue while >now >we fully unroll the loop. Missed SLP introduce memory mismatch stall. > >I am bit lost on why SLP gives up though.
I will have a look - it shouldn't give up so easily. Maybe it's a costing issue. Richard. >Honza
